pptx - UCLA

Report
Human Re-identification: A Survey
Yuanlu Xu
merayxu@gmail.com
http://vision.sysu.edu.cn/people/yuanluxu.html
Chapter 1
Introduction
Problem
In non-overlapping camera networks, matching the same individuals across
multiple cameras.
Application
Setting
S vs. S
M vs. S
M vs. M
Difficulty
Non-overlapping Camera Views
Irrelevant negative samples, difficult to train
classifiers
Difficulty
View/Pose Changes
Difficulty
Occlusions
Carried objects occlude the person appearance
Difficulty
Illumination Changes
Need illumination-invariant features or lightamending process
Difficulty
Difficulty
Large Intra-class Variations & Limited Samples
for Learning
Chapter 2
Literature Review
Literature Review
How to perform re-identification?
1. Finding corresponding local patches/parts/regions of two persons.
1. M. Farenzena et al., CVPR10 – by segmentation
2. D.S. Cheng et al., BMVC11 – by detection
3. R. Zhao et al., CVPR13 – by salience
2. Decreasing the intra-class difference caused by view, pose,
illumination changes. - Learning transformation among cameras.
1. W.S. Zheng et al., BMVC10, CVPR11, TPAMI12 – by projection basis pursuit
2. W. Li et al., ACCV12 – by selected training samples
3. W. Li et al., CVPR13 – by gating network
Features and distance measurement.
Datasets.
Literature Review: Finding Correspondence by Segmentation
Foreground Mask:
N. Jojic et al. “Component Analysis:
Modeling Spatial Correlations in Image
Class Structure”. CVPR 2009.
X-Axis of Asymmetry:
This separates regions that strongly
differ in area and places, e.g.,
head/shoulder.
M. Farenzena et al., “Person Re-Identification by Symmetry-Driven Accumulation of Local
Features”, CVPR 2010.
Literature Review: Finding Correspondence by Segmentation
X-Axis of Symmetry:
This separates regions with strongly
different appearance and similar area,
e.g., t-shirt/pants.
Y-Axis of Symmetry:
This separates regions with similar
appearance and area.
Literature Review: Finding Correspondence by Detection
Characteristics:
1. The body is decomposed into a set of
parts, their configuration  = 0 , 1 , . .  .
2. Each part  = { ,  ,  ,  }, position,
orientation and scale, respectively.
3. The distribution over configurations can
be factorized as:
and the part relation is modeled as a
Gaussian in the transformed space:
D.S. Cheng, M. Cristani, et al., “Custom pictorial structures for re-identification”, BMVC 2011.
Literature Review: Finding Correspondence by Detection
Framework:
1. Constructing pictorial structures model for every image.
2. Extracting features from each part.
3. Re-id by part-to-part comparison.
Literature Review: Finding Correspondence by Salience
Intuition:
Recognizing person identities based on
some small salient regions.
Salient regions:
1. Discriminative in making a person
standing out from their companions.
2. Reliable in finding the same person
across different views.
R. Zhao et al., "Unsupervised Salience Learning for Person Re-identification “, CVPR 2013.
Literature Review: Finding Correspondence by Salience
Adjacency Constrained Search:
Restricting the search range for each image
patch.
Similarity Score:
(, ): Euclidean distance between patch
feature x and y.
Literature Review: Finding Correspondence by Salience
Salience for person re-ID:
salient patches are those possess
uniqueness property among a specific set.
K-Nearest Neighbor Salience:
 ( (⋅)): the distance between the
patch and its k-th nearest neighbor.
The salient patches can only find limited
number of visually similar neighbors.
Literature Review: Finding Correspondence by Salience
Matching Similarity:
For each patch from image A, searching for
nearest neighbor in B:
Similarity is measured by the difference
in salience score, the similarity between two
patches:
Literature Review: Learning Transformation
x: difference vector
Distance:
Intuition:
Maximizing the probability of a pair of true match having a smaller
distance than that of a pair of related wrong match.
True match difference
False match difference
Objective Function:
W.S. Zheng et al., "Person re-identification by support vector ranking ", BMVC 2010.
W.S. Zheng et al., "Person Re-identification by Probabilistic Relative Distance Comparison", CVPR 2011.
W.S. Zheng et al., “Re-identification by Probabilistic Relative Distance Comparison”, TPAMI 2012.
Literature Review: Learning Transformation
Distance:
Intuition:
Training samples is different from
test samples. To learn the optimal
metric, weights of training samples
need to be adjusted.
Objective:
Learn generic metric 0 from whole
training set and learn adaptive
metrics  from training samples
similar to the test sample.
W. Li et al., "Human Re-identification with Transferred Metric Learning“, ACCV 2012.
Literature Review: Learning Transformation
Intuition:
1. Jointly partitions the image spaces
of two camera views into different
configurations according to the
similarity of cross-view transforms.
2. The features optimal for
recognizing identities are different
from those for clustering cross-view
transforms.
W. Li et al., "Locally Aligned Feature Transforms across Views “, CVPR 2013.
Literature Review: Learning Transformation
Gating Network:
Gating function:
Local Expert:
Literature Review: Features and Distance Measures
1. M. Farenzena et al., CVPR10.




HSV histogram, weighted by distance to Y symmetry axis.
Maximally Stable Color Regions (MSCR).
Recurrent High-Structured Patches (RHSP), replaced to LBP in the public code.
Distance:
2. D.S. Cheng et al., BMVC11.
 HSV histogram, distinct count for the full black color, different weight for different part.
 Maximally Stable Color Region(MSCR).
3. R. Zhao et al., CVPR13.
 Segmented a person image into local patches.
 Lab histogram, SIFT.
4. W.S. Zheng et al., BMVC10, CVPR11, TPAMI12.
 Divided a person image into six horizontal stripes.
 RGB, YCbCr, HSV histograms, Schmid and Gabor.
5. W. Li et al., ACCV12, CVPR13.
 Segmented a person image into local patches.
 HSV, LBP histogram, HOG and Gabor.
Literature Review: Datasets
VIPeR
 632 person, 2 images per
person.
 Occlusions.
 Pose/View changes, shot
by two cameras.
 Heavy illumination
changes.
D. Gray et al., "Viewpoint Invariant Pedestrian Recognition
with an Ensemble of Localized Features”, ECCV 2008.
Literature Review: Datasets
i-LIDS
 119 person, total 476
images.
 Occlusions, conjunctions.
 Pose/View changes, shot
by multiple cameras.
 Moderate illumination
changes.
W.S. Zheng et al., "Person Re-identification by Probabilistic
Relative Distance Comparison", CVPR 2011.
Literature Review: Datasets
ETHZ
 146 person, total 8555
images.
 Occlusions, conjunctions.
 No view changes.
A. Ess et al., "Depth and Appearance for Mobile Scene
Analysis", ICCV 2007.
Literature Review: Datasets
CAVIAR4REID
 72 person, total 1220
images.
 Scale changes.
 Limited view changes.
D.S. Cheng et al., "Custom Pictorial Structures for Reidentification", BMVC 2011.
Literature Review: Datasets
EPFL
 30 person, total 70 reference
images, 294 query images
and 80 group shots.
 Limited occlusions and
conjunctions.
 multiple view/pose changes.
 Limited illumination changes.
Y. Xu et al., "Human Re-identification by Matching Compositional
Template with Cluster Sampling", ICCV 2011, under review.
Literature Review: Datasets
CAMPUS-Human
 74 person, 5 reference
images per person, total
1519 query images and
213 group shots.
 Occlusions, conjunctions.
 multiple view/pose
changes.
 Limited illumination
changes.
Y. Xu et al., "Human Re-identification by Matching Compositional
Template with Cluster Sampling", ICCV 2011, under review.
Vista
How about something more challenging?
Retrieving a character in music
videos.
Questions?
Literature Review: Features and Distance Measures
Feature:
1.
2.
3.
Weighted Color Histogram.
- HSV histogram
- weighted by distance to Y
symmetry axis.
Maximally Stable Color
Regions (MSCR).
Recurrent High-Structured
Patches (RHSP).
- replaced to LBP in the
public code.
Distance:
M. Farenzena et al., “Person Re-Identification by Symmetry-Driven Accumulation of Local
Features”, CVPR 2010.
Literature Review: Features and Distance Measures
Framework:
1.
2.
3.
Constructing pictorial model for every image.
Extracting features from each part.
1. Maximally Stable Color Region(MSCR).
2. HSV histogram. - Distinct count for the full black color, different
weight for different part.
Re-id by part-to-part comparison.
D.S. Cheng, M. Cristani, et al., “Custom pictorial structures for re-identification”, BMVC 2011.
Literature Review: Learning Transformation
Objective Function:
Constraining the distance matrix M:
W.S. Zheng et al., "Person re-identification by support vector ranking ", BMVC 2010.
W.S. Zheng et al., "Person Re-identification by Probabilistic Relative Distance Comparison", CVPR 2011.
W.S. Zheng et al., “Re-identification by Probabilistic Relative Distance Comparison”, TPAMI 2012.
Literature Review: Learning Transformation
Objective Function:
Weight of training sample
the distance between samples
of different persons.
the distance between
samples of the same person.
W. Li et al., "Human Re-identification with Transferred Metric Learning“, ACCV 2012.
Literature Review: Learning Transformation
Gating Network:
Parameters:
1. ( ,  ) is the gating parameters to
partition samples into different
configurations.
2. ( ,  ) is the alignment matrix pair
for expert k, which projecting the two
samples into a common feature space.
W. Li et al., "Locally Aligned Feature Transforms across Views “, CVPR 2013.

similar documents